Scholarlink

<p>Reliable health monitoring of aerospace vehicles during long-term storage is hindered by environmental coupling and severe measurement noise, causing physical inconsistencies in pure data-driven models. This paper proposes the Physics-Consistent Contrastive Multimodal Alignment Framework (PC-PIMAT), a compact 1.42-million parameter network for robust state estimation. To suppress telemetry anomalies, a Physics-Gated Unit (PGU) incorporates the first-order derivative of material aging kinetics as a low-pass filter to calibrate cross-modal attention. Concurrently, a physics-consistent contrastive regularization builds temporally validated pairs to align multi-sensor features and physical embeddings within a bounded latent manifold. Validation on a 15-year storage dataset under five independent random seeds demonstrates that PC-PIMAT cuts prediction error by 56.3% over baselines and achieves a 96.5% parameter ranking accuracy. Post-hoc causal feature perturbations confirm accurate isolation of dominant thermal and hygro-thermal degradation drivers. Furthermore, the framework maintains stable state estimation under extreme non-Gaussian corruptions, including sensor bias drift and transient burst spikes, with a noise standard deviation up to 10. The proposed architecture effectively bridges the simulation-to-real (Sim-to-Real), offering a robust and interpretable asset lifecycle monitoring methodology under intense measurement uncertainty.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A physics-consistent intelligent framework with contrastive multimodal alignment for robust state estimation in complex aircraft storage environments

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月21日

查看全文
  • Yao Lei,
  • Jianyin Zhao,
  • Weimin Lv,
  • Song Gao

摘要

Reliable health monitoring of aerospace vehicles during long-term storage is hindered by environmental coupling and severe measurement noise, causing physical inconsistencies in pure data-driven models. This paper proposes the Physics-Consistent Contrastive Multimodal Alignment Framework (PC-PIMAT), a compact 1.42-million parameter network for robust state estimation. To suppress telemetry anomalies, a Physics-Gated Unit (PGU) incorporates the first-order derivative of material aging kinetics as a low-pass filter to calibrate cross-modal attention. Concurrently, a physics-consistent contrastive regularization builds temporally validated pairs to align multi-sensor features and physical embeddings within a bounded latent manifold. Validation on a 15-year storage dataset under five independent random seeds demonstrates that PC-PIMAT cuts prediction error by 56.3% over baselines and achieves a 96.5% parameter ranking accuracy. Post-hoc causal feature perturbations confirm accurate isolation of dominant thermal and hygro-thermal degradation drivers. Furthermore, the framework maintains stable state estimation under extreme non-Gaussian corruptions, including sensor bias drift and transient burst spikes, with a noise standard deviation up to 10. The proposed architecture effectively bridges the simulation-to-real (Sim-to-Real), offering a robust and interpretable asset lifecycle monitoring methodology under intense measurement uncertainty.